For two years, agentic commerce ran on demos and projections. On October 1, 2026, it acquired something better: a measurement baseline. Digiday’s weekly data column assembled five independent research fleets into a single frame, and each one lit up a different organ of the same organism. HUMAN Security weighed the traffic. Adobe priced the visits. Jellyfish and Tinuiti mapped how assistants choose. Dentsu polled the marketers now lobbying the machine. The same week, Salesforce published its 2026 holiday predictions from more than 1.5 billion shopper signals, and PYMNTS published a playbook for the eight-protocol stack merchants are being told to support.
Put together, they answer four questions with real numbers. How much agent traffic is there, really. What is an AI-referred shopper worth. How do the assistants decide what to show. And what is the marketing world doing about it. All four answers point the same direction: the gatekeeper is real, it is heavy, and it is already deciding.
One question is missing from every dataset. What does the gatekeeper read, and is any of it true.
The Volume: Muse and the 72 Percent
Start with the traffic, because the traffic settled the “is this happening” debate. HUMAN Security, the bot-management firm whose sensors sit inside a large share of major websites, began tracking Meta’s Muse agent on September 22. In under two days it logged 5.6 million requests from Muse. Reaching that same request volume took ChatGPT eleven months.
Since then, Muse has averaged 72 percent of the daily AI-agent requests HUMAN observes across its network. The composition matters more than the total: 84 percent of Muse requests target product and search pages, and in a recent sample Muse was 25 percent more likely than other agents to reach checkout. This is not ambient model traffic touching ecommerce by accident. It is shopping-shaped traffic, aimed at shelves, that converts. HUMAN’s own read is that Muse may be the first mass-market AI agent, helped by a generous free tier and distribution across Facebook, Instagram, Messenger and WhatsApp.
Remember what this means for the corpus. Every one of those product-page and search-page visits is a machine reading the same surfaces humans read: listings, specs, and above all reviews. At 84 percent product-page concentration, the review corpus is no longer background context for agent traffic. It is the input layer.
The Economics: The Flip Adobe Caught
The second dataset answers the question retailers actually ask: is this traffic worth anything. Adobe, drawing on more than one trillion visits to US retail sites, found that shoppers arriving from AI referrals generated 53 percent more revenue per visit than non-AI traffic in July 2026. They added to cart 28 percent more often and converted 60 percent higher.
The most important number is the comparison. Twelve months earlier, non-AI visits were worth 128 percent more than AI visits. Adobe’s Q3 AI Traffic Trends Report calls it plainly: the conversion gap in retail has disappeared. Within a single year, AI-referred traffic went from the worst-performing channel on the site to the best, a swing of more than 180 points of relative value.
Part of this is selection: someone who asks a chatbot for a product is further along than someone idly browsing. Adobe itself credits assistants getting better at surfacing relevant products to higher-intent shoppers. But the macro context makes the flip ominous rather than merely encouraging. Salesforce’s Shopping Index shows global digital traffic grew 18 percent in Q2 while order volume grew just 1 percent, with cart abandonment at 82 percent. Attention is abundant and orders are scarce. AI referrals are one of the only channels converting attention efficiently, which means retailers are becoming dependent on the gatekeeper’s judgment at exactly the moment the gatekeeper’s judgment becomes the product.
Salesforce’s own holiday forecast sizes the dependency: 20 percent of all 2026 holiday ecommerce traffic will originate from AI chat agents, a blend of consumer-facing bots, autonomous backend agents, and competitor scrapers feeding price-matching algorithms. Consumer behavior has already moved ahead of the forecast: 50 percent of shoppers report using an AI assistant somewhere in their buying journey, up 67 percent year over year, reliance on AI assistants as the first stop in shopping grew 200 percent from May 2025 to May 2026, and use of traditional search engines and marketplaces each fell 15 percent. Even the physical store is in on it: 12 percent of in-store shoppers now turn to an AI assistant at the shelf.
The Menu: How the Gatekeeper Chooses
If agents are now routing a fifth of holiday traffic, the obvious next question is what determines which products they route it to. Two datasets published this week give the first honest map, and the map is balkanized.
Jellyfish’s Share of Model tool tracked product recommendations from three assistants across eight product categories in the US, UK, Australia and Singapore. Asked for toy recommendations, Alexa for Shopping surfaced 177 brands, and nearly every one came from Amazon’s own store. The identical request put to ChatGPT produced recommendations spanning 24 retailers. Google’s AI Mode went to 37. Three assistants, three completely different shelves.
Tinuiti measured the citation side of the same phenomenon, tracking which ecommerce sites the major assistants reference in their answers. Microsoft’s Copilot cites retailers the most: in July, the top 100 retailers accounted for 34 percent of Copilot’s citations, more than triple their share on Google’s AI Mode and more than eight times their share on ChatGPT. One infrastructure quirk explains much of it: Reddit blocks Copilot’s crawlers, so the forum layer of Copilot’s diet is missing and stores fill the space. Across all six assistants Tinuiti tracks, retailers’ combined share of citations slipped to 14 percent in July.
Three conclusions follow. First, there is no single AI shelf: visibility is a property of each assistant’s corpus, its crawl permissions, and its commercial wiring. Second, when the gatekeeper is owned by the store, the shelf narrows to one seller. Amazon’s September 20 blockade of Muse and Alexa’s 177-brand toy aisle that is almost entirely Amazon are not two facts about the market. They are one strategy observed from two sides: control what the agent can reach, and the agent’s choice collapses into your inventory. Third, and least examined: what the assistants cite when they do not cite retailers is forums, review platforms, and publisher roundups, the exact surfaces where fake-review operations, covered here last week, plant manufactured consensus. The citation diet and the contamination zone are the same place.
The Lobby: Marketers Optimize to Be Picked
The fourth dataset measures the response. Dentsu Creative’s latest CMO survey, polling 1,950 senior marketers across 14 markets, found 59 percent are already optimizing content and SEO for AI-enabled search, tied for the top innovation strategy alongside social commerce and AI-driven personalization, and half are investing in agentic commerce for customer experience.
Salesforce’s recommendation to merchants drops the feudal framing directly: treat AI platforms as your most critical B2B data syndication partners. Product feed syndication, structured APIs, optimized product graphs delivering real-time pricing, exact specs, live inventory. In Salesforce’s words, “if your product data isn’t integrated into the feeds powering these agents, your inventory won’t be part of the consideration set.”
The protocol stack being assembled for exactly this syndication now runs eight deep, as PYMNTS’s standards playbook lays out: Anthropic’s MCP for tools and data, Google’s A2A for agent-to-agent communication, OpenAI and Stripe’s ACP for discovery and checkout, Google and Shopify’s UCP across the commerce workflow, Google’s AP2 with cryptographically signed purchase mandates, Visa’s Trusted Agent Protocol and Mastercard’s Agent Pay for agent identity and payment, and Stripe and Tempo’s Machine Payments Protocol for machine-to-machine settlement. Mastercard’s Sabrina Tharani frames the shift precisely: “We don’t see [agentic] so much as a new channel, as much as a new interface for interacting and engaging with commerce.”
Read the whole stack carefully and notice what every layer carries: identity, authorization, price, spec, stock, settlement. The syndication contract the industry is converging on moves accurate commerce facts about the product’s logistics and nothing about the product’s merit. There is no field for whether the 4.7 stars behind the listing survive contact with fake-review detection. The lobby is spending 59-percent-of-the-CMO-corps budgets to be findable by machines, and zero protocol surface exists to be verifiable to them.
The Part Nobody Measured: What the Gatekeeper Believes
Here is the finding that should end the debate about whether an evidence layer is a nice-to-have. Salesforce reports that 74 percent of shoppers now trust the product recommendations they receive from AI chat. And on brand-owned sites the number turns sharper: 41 percent of shoppers say a brand AI assistant answering their questions makes them “much more confident” in a purchase, to the point where, in Salesforce’s own phrasing, “it replaces the need to read reviews entirely.” Another 36 percent are somewhat more confident. Only 5 percent distrust the answers.
Sit with the mechanics of that sentence. The review corpus was the last widely-used epistemic check in ecommerce, the one instrument shoppers operated themselves, imperfectly, against sponsored listings and seeded ratings. Consumers are now deleting it from their own process, willingly, because a confident assistant answered their questions. And what does the assistant answer from? The same corpus, scraped, cached, and averaged, minus the skepticism.
This is the compounding failure this column has tracked all month, now with a consumer-behavior number attached. The contamination is documented: Amazon itself blocked more than 275 million suspected fake reviews in 2024, the FTC’s Consumer Review Rule carries civil penalties above $53,000 per violation, and Singapore’s Competition and Consumer Commission spent last week detailing Reputifly, an operation with a rating calculator that computed how many five-star reviews a business needed to buy, packages at S$219 for 35 reviews, generative AI that staged initial reservations to mimic authentic doubt, and a warranty to replace up to 30 percent of any review a platform caught. The platform layer routes a fifth of holiday traffic through product pages. The consumer layer stops reading the reviews. The corpus layer is provably for sale. Each dataset published this week strengthens one leg of that tripod, and none of them addresses the joint.
The obvious objection is that agents will get better at spotting fakes themselves. Nothing in this week’s data supports that hope, and the citation evidence points the other way: assistants build product beliefs from whatever their crawlers can reach, and the crawlable web is the manipulated web. A model does not skim reviews the way a tired human does at midnight, applying half-remembered folk heuristics about verified purchases. It ingests them as evidence, aggregates them, and emits a recommendation whose fluency launders whatever was in the input. For an agent, a purchased five-star rating is not noise to be discounted. It is a signal to be obeyed.
Evidence as a Feed: The Missing Row in the Syndication Table
If Salesforce is right that AI platforms are now B2B data syndication partners, then the syndication contract is incomplete in a specific, fixable way. Merchants are being told to ship price, spec, and inventory to every protocol in the alphabet soup. Nobody is shipping trust. The complete feed has four rows the current stack lacks.
Filtered review data, before aggregation. Any average computed over an unfiltered corpus is a precise answer to the wrong question. Fake, incentivized, and low-information reviews need to be removed before a score is computed, not explained away after. GoBuy’s Smart Score, 0 to 100, is computed only after that filtering.
Quality-weighted scores, not volume-weighted stars. Ten thousand reviews of unknown provenance should not outrank nine hundred verified ones. Weighting by review quality, not review count, breaks the economics of the rating calculator: buying volume stops working when volume is not the input.
Persistence, not snapshots. A reputation campaign can spike a score for a week. The GoBuy Verified badge requires a filtered score above 80 held for 90 days, exactly the property a deadline-driven burst of purchased reviews cannot manufacture.
Machine-native delivery. None of this matters if it cannot enter the agent’s context. GoBuy exposes product evidence over MCP at gobuy.ai/api/mcp, so any shopping agent can consult verified review intelligence as a tool call before it recommends or buys, the same pattern the security and intelligence industries adopted this week with their own MCP servers. For the human side of the ledger, the Chrome extension injects the trust panel directly onto Amazon pages, so the person and the machine finally read the same evidence.
The five datasets settled the volume, the value, the menu, and the lobby. The gatekeeper is here, it is worth 53 percent more per visit than the traffic you already optimize for, and it decides from a reading list nobody has audited. The next dataset the industry needs is not another measurement of how agents choose. It is a measurement of whether what they read is true. That is the dataset we ship every day. Check products before you buy at gobuy.ai, and if you build agents, wire them to the evidence layer at gobuy.ai/agent-docs.